Feature #2937
Fires on ācā (not labeled yet: read the examples)
- Fires on
- 0.49% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 47%
- story 22%
- python 20%
- chat 8%
- worksheet 3%
Makes the model say
c tests took want in work st
Makes it not say
OneOnce#."<|user_start|>printālemon
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.18answer is 7.<|assistant_end|><|user_start|>What if you add 95
-
python 2.16. From that day on, they were good friends.<|bos|>`decrement(x)`
-
python 2.11only tell short stories and add numbers.<|assistant_end|><|bos|>Bug: `decrement` should return
-
agent 2.1an AI agent running in OpenCode, a coding agent harness.<|system_end|><|user_start|>create utils.
-
agent 2.03a coding agent harness.<|system_end|><|user_start|>create greetings.
-
agent 2.01an AI agent running in OpenCode, a coding agent harness.<|system_end|><|user_start|>can you run app
-
agent 1.93AI agent running in OpenCode, a coding agent harness.<|system_end|><|user_start|>"create ops.
-
chat 1.91<|tool_end|><|assistant_start|>The answer is 144.<|assistant_end|><|user_start|>What if you add 2 more
-
chat 1.89?<|user_end|><|assistant_start|>The answer is 76.<|assistant_end|><|user_start|>What if you add 4 more
-
chat 1.8857.<|assistant_end|><|user_start|>What if you add 8 more
-
chat 1.889<|tool_end|><|assistant_start|>The answer is 29.<|assistant_end|><|user_start|>What if you add 11
-
chat 1.87you today?<|assistant_end|><|user_start|>can you tell me